Landslide susceptibility prediction based on image semantic segmentation

The visual characteristics of landslide susceptibility have not yet been fully explored. Professional or trained technicians have to take much time and effort to interpret remote sensing images and locate landslides accordingly. Although conventional machine learning methods based on hand-crafted features for landslide susceptibility prediction (LSP) have acquired remarkable performance, they have certain requirements for prior knowledge. Aiming to learn complex and inherent visual patterns of landslides through minimal manual intervention and achieve fine-grained prediction, in this paper, we define LSP as a semantic segmentation problem on optical remote sensing images. Six widely used semantic segmentation models including Fully Convolutional Network, U-Net, Pyramid Scene Parsing Network, Global Convolutional Network (GCN), DeepLab v3 and DeepLab v3+ are introduced and evaluated for LSP. As the lack of landslide datasets, an open labeled landslide dataset of remote sensing imagery is created for research. The results show that GCN and DeepLab v3 are more applicable for this problem scenario, and the best Mean Intersection-over-Union and Pixel Accuracy of models are 54.2% and 74.0% respectively, which could be further improved by more targeted network architectures. In conclusion, semantic segmentation methods are demonstrated to be effctive for predicting new potential landslides based on remote sensing images.